Does high school homework increase academic achievement?

805 high school graduates (438 female, 367 male)

US adolescents in the PSID Child Development Supplement and Transition to Adulthood, followed to age 20

Education Economics 25(1), 45-59; full text read from IZA discussion paper 8142 (April 2014). The distinctive feature is measurement: homework comes from TIME DIARIES for weekday and weekend days rather than retrospective questionnaires, which removes the self-report inflation that Cooper's reporter analysis shows drives the correlational literature. Outcomes are high school GPA (a teacher-assigned measure, self- or parent-reported on a 0-100 scale) and college attendance by age 20 - there is no independent standardized test anywhere in this paper. Estimation is OLS and probit, then jointly with a homework equation by limited-information maximum likelihood using exclusion restrictions: for girls, whether the diary day was a Friday, whether it fell in spring, and county daily precipitation; for boys, Friday and Saturday diary days and being older than the state minimum driving age.

The uncorrected estimates show a male-only association: 'we find a significant positive relationship between homework and college attendance by age 20 for males, regardless of homework measure... An extra hour of total homework per night (i.e. 7 hours per week) would increase the probability of attendance by nearly 5 percentage points', and about 2.5 percentage points on a 100-point GPA scale. Nothing for females on either outcome. THE ENDOGENEITY-CORRECTED ESTIMATES ARE NULL: 'we find that homework is no longer a significant predictor of college attendance by age 20 for males', and a Tobit robustness check finds 'no significant effects of homework on academic achievement'. The authors prefer the uncorrected specification on the grounds that Wooldridge tests fail to reject exogeneity in the GPA equations and 'the estimated correlation coefficients between the error terms across equations are never significant at the 5% level'. That is a defensible position, but the honest summary for an archive is that every estimate in this paper which attempts to handle selection is a null.

High in the preferred specification and only partially addressed in the corrected one. Homework time is a behaviour, and the outcomes are GPA and college-going, both of which are predicted by the same conscientiousness and ability that predict doing homework. The instruments - rain, day of the week, driving age - are plausibly exogenous but weak, and the paper's own diagnostics cannot distinguish 'exogeneity holds' from 'the instruments have no power'.

Unreplicated. It fits a pattern that recurs across this topic: Eren and Henderson's homework coefficient falls 73% when one lagged test score is added, Rønning's falls from -0.12 SD to essentially zero when individual covariates are added, and Fernández-Alonso's flips negative when prior grades are added. Here the estimate does not shrink so much as vanish when the selection equation is estimated jointly.